Vals AI's MysteryMechanism Benchmark Tests Scientific Rediscovery

Vals AI released MysteryMechanism, a benchmark meant to test scientific discovery rather than recall of already-published results. Vals AI's post explains the difficulty with grading scientific discovery directly: new results are hard to verify, which makes them hard to benchmark. MysteryMechanism works around that by sealing a mathematical mechanism the model must rediscover through its own bounded experiments, rather than asking a question with a known answer, so what gets scored is the model's process of forming and testing a hypothesis. Vals AI reports the benchmark separates frontier models sharply, with GPT-6 Astra scoring roughly 20 percentage points above GPT-5.6 Sol. The chart accompanying the post, which likely shows a fuller model ranking, was not reviewed here, so how other tested models compare beyond that one reported gap is not established.